基于双流自适应图卷积网络的管制员睡岗行为识别

Sleeping on duty behavior recognition of air traffic controller based on two-stream adaptive graph convolution network

为识别空中交通管制员的睡岗行为,减少管制差错,保障航空器飞行安全,提出了一种基于双流自适应图卷积网络的管制员睡岗行为识别方法。该方法设计双流网络分别处理管制员骨架的一阶信息和二阶信息,实现对骨架数据的充分提取;通过自适应学习的骨骼拓扑连接矩阵,挖掘管制员不同关节之间的功能连接关系;同时在卷积层引入时空通道注意力机制,增强管制员睡岗行为识别模型在时间、空间、通道3个方向提取重要信息的能力。仿真结果表明,该方法能有效识别管制员3种睡岗行为,相较于传统的时空图卷积网络,识别准确率提高了3.08百分点,达到95.03%,可以提高民航运行安全管理水平。

A method for recognizing the sleeping-on-duty behavior of air traffic controllers based on a two-stream adaptive graph convolutional network is proposed to reduce control errors,ensure flight safety,and enhance the ability of civil aviation operation safety management.Firstly,the OpenPose algorithm is utilized to extract the skeleton key points of the controller in each frame of the video.Then,the matching optimization and clustering algorithms are applied to cluster and combine the key points to obtain the skeleton joint graph of the controller,and the same joint points of different frames are connected to obtain the controller skeletal spatial-temporal graph.The controller skeletal spatial-temporal graph is then transformed into joint flow data and bone flow data,and the two-stream network is used to process the joint flow information and bone flow information respectively,achieving full extraction of the first-order and second-order information of the skeleton data.Next,the adaptive learning bone topology connection matrix is employed to explore the functional connection between different joints of the controller,while the STC attention mechanism is introduced in the convolutional layer to enhance the ability of the sleeping-on-duty behavior recognition model to extract important information in the time,space,and channel dimensions.Finally,the Controller Working Status Dataset is collected for model training,and the recognition of three sleeping on duty behavior,nodding off,falling asleep with head down,and falling asleep with head tilted back,is achieved on the test set,verifying the effectiveness of the adaptive graph convolutional network and two-stream network design.Experimental results show that compared to the spatial-temporal graph convolutional network.The proposed method achieves a more significant improvement in the recognition accuracy of sleeping on duty behavior,with a recognition accuracy of 95.03%,which can strengthen the effective control of the working behavior of controllers and enhance the ability of civil aviation operation safety management.